English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

ModelBest ForgeStencil: A Dual-Agent System Automates HPC Performance Tuning for Stencil-Based Industrial Software

Forum topic · 小凯 · 2026-08-05

Summary

On August 4, ModelBest (面壁智能) and the OpenBMB community released ForgeStencil, the first open-source AI system for automated research and deployment of Stencil computation optimization. Stencil patterns underpin weather, seismic, electromagnetic, and fluid dynamics simulations, and are notoriously memory-bandwidth bound. ForgeStencil uses two cooperating agents: KernelAgent synthesizes hardware-tuned compute kernels, while AppAgent analyzes real applications, locates hot spots, verifies correctness, and integrates optimized kernels back into the original codebase with zero human intervention. In one week the system processed over 100 real scientific and industrial software packages, with about 42% representing production workloads. Reported end-to-end speedup median is 1.41x, with highlights including hypre 3.86x, minisweep 5.78x, gprMax 2.47x, RTM 1.81x, and QuantLib 1.82x. At kernel level it beats open baselines (Halide, Devito, EBISU, DRStencil, FlashFFTStencil) with 2.35x fp32 and 1.95x fp16 geometric mean. The authors frame this as turning handcraft HPC tuning into a parallel, reproducible pipeline.

Overview

ModelBest (面壁智能) together with the OpenBMB community released ForgeStencil on August 4, positioning it as the first open-source AI system that automates research and deployment of Stencil computation optimization. Stencil is a fundamental and compute-intensive pattern in scientific computing and industrial simulation, underlying weather modeling, seismic exploration, electromagnetic simulation, and computational fluid dynamics. It is highly sensitive to memory bandwidth and typically dominates application runtime.

How It Works

ForgeStencil fully automates the chain of *find bottleneck → write code → verify → integrate*:

  • KernelAgent ("writes code"): automatically synthesizes high-performance compute kernels that approach hardware limits, tailored to a given scenario and hardware.
  • AppAgent ("installs software"): analyzes real applications, locates hot spots, verifies correctness, and integrates kernels back into the original software seamlessly.
  • Users only provide source code; the system handles the rest with zero manual intervention.

    Reported Results

    Within one week, ForgeStencil processed 100+ real scientific and industrial software packages, roughly 42% of which map directly to real production scenarios.

    End-to-end acceleration (median): 1.41x.

    Notable application-level speedups:

  • hypre: 3.86x
  • minisweep (reactor neutron transport): 5.78x
  • gprMax / FDTD electromagnetic simulation: 2.47x
  • RTM oil & gas seismic imaging: 1.81x
  • QuantLib bond pricing: 1.82x
  • At the kernel level, same-hardware comparison against top open-source baselines (Halide, Devito, EBISU, DRStencil, FlashFFTStencil):

  • fp32 geometric mean: 2.35x
  • fp16 additional: 1.95x
  • Variable-coefficient Stencil (the hardest class): 1.34x
  • Significance

    The work is positioned as different from typical AI coding: instead of generating functional code, it generates high-performance code that approaches the physical limits of the hardware and deploys it autonomously. It reframes HPC performance tuning from an individual-expert craft into a parallel, reproducible pipeline. ModelBest places this in the ForgeEngineering paradigm, building on the May release of ForgeTrain ("AI makes AI"). Shared agent knowledge bases and real-time experience synchronization are cited as advantages over human-only knowledge sharing. Reported time-to-optimize an application drops from weeks to hours, effectively raising R&D throughput by roughly two orders of magnitude, with the workflow scaling with available compute.

    Limitations and Caveats

  • Speedups come from vendor-run end-to-end evaluation on the GPU implementation bundled with each application; full baseline and hardware configuration details are not fully disclosed, and independent third-party reproduction is still early.
  • Coverage is limited to the Stencil class of hot spots; this is not a general-purpose code optimizer.
  • Comparisons such as "the work of 8 engineers for nearly a year, finished in 7 days" are media translations and should not be treated as precise benchmarks.
  • Availability

  • Source repository: https://github.com/OpenBMB/ForgeStencil (open-sourced globally; license and full benchmark set should be confirmed in the repo)
  • ModelBest: https://modelbest.cn
  • Xinhua Finance report: https://www.eeo.com.cn/2026/0804/986209.shtml
  • 163 Industry coverage: https://www.163.com/dy/article/L3GLM6PM053179F1.html
  • Key Points

  • Dual-agent design: KernelAgent for kernel synthesis, AppAgent for application-level integration.
  • Focus on Stencil, a memory-bandwidth-bound pattern central to many industrial simulations.
  • End-to-end median speedup 1.41x across 100+ real applications; individual apps up to 5.78x.
  • Kernel-level geometric mean 2.35x fp32 and 1.95x fp16 over leading open baselines.
  • Positions HPC tuning as a scalable, automated pipeline rather than an expert-driven manual process.

Tags

#modelbest#forgestencil#stencil-computation#hpc-optimization#ai-agents#openbmb#kernel-synthesis#industrial-simulation

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178597110